Datasets:

Bc-AI
/

SFT-Ultra

license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - reasoning
  - sft
  - distillation
  - merged
  - chain-of-thought
size_categories:
  - 1M<n<10M

Mega Merged Reasoning SFT Dataset

Auto-merged from 44 source datasets → 3,681,880 rows.

Schema

Every row is:

{
  "messages": [
    {"role": "system",    "content": "..."},
    {"role": "user",      "content": "..."},
    {"role": "assistant", "content": "<think>\n...\n</think>\n\nFinal answer..."}
  ],
  "thinking": "extracted chain-of-thought (or null)",
  "source":   "org/dataset-name"
}

messages — standard OpenAI / TRL chat format
thinking — isolated CoT when available (also embedded inline in <think> tags)
source — which original repo this row came from

Formats auto-detected

#	Strategy	Description
1	`messages` list	OpenAI-style `role`/`content` dicts
2	`conversations` list	ShareGPT-style `from`/`value` dicts
3	`conversation` list	Variant key name
4	`chat`/`turns`/`dialog` list	Other conversation keys
5	`messages` with `reasoning_content`	OpenMed / R1 style per-turn reasoning
6	`prompt` + `response`	Flat pair
7	`instruction` + `output`	Alpaca-style flat pair
8	`question` + `answer`	QA-style flat pair
9	`query` + `completion`	Completion-style flat pair
10	`problem` + `solution`	Math-style flat pair
11	`input` + `output`	Generic flat pair
12	Any flat pair + `thinking`/`cot`/`reasoning`	Separate CoT column
13	`complex_cot` + `response`	Medical-o1 style
14	Nested `provenance.prompt`	Crownelius coding datasets
15	`system` / `system_prompt` field	System message extraction
16	ChatML raw text	`<
17	Alpaca raw text	`### Instruction:` parsing
18	Inline `<think>...</think>`	Extracted + preserved
19	JSON-string conversation	Stringified list parsing
20	Single `text` blob	Last-resort extraction
21	Brute-force string scan	Two longest strings → user/assistant
22	Prompt-only (no response)	Kept as user-only row

Mega Merged Reasoning SFT Dataset

Schema

Formats auto-detected

Source datasets